Here's a comparison I keep running into.
Business A has ChatGPT everywhere. The team drafts emails with it, the owner asks it strategy questions, someone's built a custom GPT for proposals. Ask them if they're "doing AI" and the answer is an enthusiastic yes.
Business B barely touches AI. But it has seven years of clean, exportable transaction history, a customer list actually linked to what those customers bought, and one set of numbers everyone agrees is the truth.
Which one is better positioned to benefit from AI over the next three years? It's B. It isn't close.
AI is a multiplier, and multipliers don't care what they multiply
Every genuinely useful thing AI can do for a small business — spot a margin trend early, flag the customers going quiet before they leave, forecast next month's demand, answer "why was March bad?" in plain English — depends entirely on the raw material underneath it.
Give those capabilities seven years of clean history and they compound. Give them scattered exports, inconsistent product names, duplicates and gaps, and they produce confident-sounding answers to questions the data can't actually support. Which is worse than no answer at all, because it feels like insight.
This is where "we should do something with AI" projects quietly die. Not at the AI stage. At the export stage — the day someone tries to pull the data and discovers it's locked in a vendor's dashboard, or that "revenue" means three different things depending on who you ask.
The unglamorous audit that predicts everything
Forget maturity models. Four questions tell you most of what matters.
How far back does your usable history go — in a form you could export today? Not "the data exists somewhere." Exportable, by you, this afternoon. Three years of transactions is worth more than any tool you could buy.
History is the one asset that can't be built retroactively.
If you exported your sales data right now, how much cleaning would it need? Try it. Genuinely — export last quarter and look. Inconsistent naming, things recorded differently over time, mystery categories. An hour of tidying is fine. A lost day is a project. "We'd need the developer" is a red flag.
Can you get at your own data without asking anyone's permission? If key numbers live only inside a vendor system, your accountant's software, or one employee's head, you don't fully own them. You rent them.
Is there one version of the numbers everyone trusts? If different people quote different figures for the same thing, every downstream tool inherits the argument. Dashboards built on contested numbers just get distrusted faster.
The good news buried in this
If you scored badly on those four, here's what I'd want you to hear: this is the cheapest problem you'll ever fix, and it's the one that makes everything else work.
Agreeing what "revenue" includes is a one-page document. Checking your exports is an afternoon. Consolidating a customer list is a small, boring project. None of it requires AI, new software, or a consultant — and every pound you later spend on tools returns multiples more because of it.
Meanwhile, the business with ChatGPT everywhere and chaos underneath hasn't built an advantage. It's built a habit of asking good questions of bad data.
Your transaction history is quietly becoming one of the most valuable assets on your balance sheet — it just doesn't appear there. When did you last check you could actually get at it?
Delphi Decide starts most engagements exactly here: finding out what your data can actually support before anyone buys anything. If you'd like to talk it through — get in touch.
